DeepCorrect: Correcting DNN Models Against Image Distortions.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep neural networks (DNNs) achieve high performance in computer vision tasks.
• Image distortions like blur and noise degrade DNN performance on real-world data.
• Networks trained on clean images are not robust to distorted inputs.

Purpose of the Study:

• To evaluate the impact of image distortions on convolutional filter activations.
• To develop a method for identifying and correcting distortion-susceptible filters in DNNs.
• To enhance the robustness of DNNs against image distortions.

Main Methods:

• Assessed the effect of Gaussian blur and additive noise on pre-trained convolutional filter activations.
• Proposed a metric to identify filters most vulnerable to noise.
• Developed DeepCorrect, which retrains specific filters using residual connections to correct activations.
• Kept other pre-trained filter outputs unchanged during retraining.

Main Results:

• Identified and ranked convolutional filters based on their susceptibility to distortions.
• DeepCorrect significantly improved DNN robustness against distorted images.
• Outperformed alternative approaches on benchmark datasets (ImageNet, Caltech-101/256, SUN-397).

Conclusions:

• DNNs are sensitive to image distortions, impacting real-world applications.
• DeepCorrect effectively enhances DNN robustness by selectively retraining filters.
• The proposed method offers a significant improvement for reliable computer vision systems facing noisy data.